Prediction Of Urban Human Flow Based On Deep Learning

With the rapid development of Chinese cities in recent years,the problem of big city disease is becoming more and more prominent.Traffic congestion is one of the most noticeable aspects.In order to alleviate traffic congestion,the government has taken many measures to improve the travel experience of people in peak period,among which the intelligent management of urban traffic is a good focus.Good intelligent management can not only optimize traffic congestion,but also add a new layer of security to people’s travel safety.In order to achieve this,it is necessary to predict the huge human flow in each area of the city every day.However,due to the changeable urban environment,it is often affected by many factors,such as the traffic trend nearby,the current emergencies,and the unpredictable weather.In order to accurately predict the inflow and outflow of people flow in each region of the city,the paper mainly does the following specific work:1.A new method based on deep learning is proposed in this paper.Its name is ST-LSTM-ResNet(Spatio-Temporal LongShort-Term Memory Residual Networks),and it is trained and tested by taxi travel data in Beijing.In order to get more accurate prediction results,in the existing St RESNET,a new end-to-end ST-LSTM-ResNet is designed according to the unique characteristics of each kind of spatiotemporal data.Specifically,the paper uses the residual network framework improved by LSTM to model and predict the future from the perspectives of periodicity,correlation and long-term trend change of human traffic data.Due to the rich and complex spatial information implied in the traffic data,each branch model is added with a separate residual network unit.In the model training,ST-LSTM-ResNet will dynamically adjust the weights of three networks according to the actual situation of the data,and simultaneously study the three networks,adjust the local parameters of each network in the three networks,and fuse them to get a comprehensive output.2.Considering the influence of temporary objective factors,the paper creatively takes air pollution factors into account.Human traffic data is easily affected by many additional objective factors,which often play a decisive role in human traffic,so they are separated for feature extraction.There are two common objective factors.The first is the change of weather,such as thunder and rain.Besides,the severity of air pollution is also an important factor.In order to extract these additional factors,the paper first obtains the weather data within the specified time,at the same time,considering that the travel also receives the influence of weather forecast data,the paper also grabs the weather forecast data.The second is whether the impact of the weekend.It can be predicted that the travel data of people on weekdays and weekends will change greatly.This feature can significantly affect the traffic volume,so the paper adjusts the final output according to whether it is weekends.3.In this paper,the micro blog data is included in the feature extraction.Considering that people’s willingness to travel is determined by both subjective and objective factors,it is very important to get the characteristics that can reflect people’s willingness to travel.Therefore,the paper grabs the data in cyberspace and uses it to predict the urban traffic.Sina Weibo is the largest Weibo platform in the era of mobile Internet.This paper crawls the keywords of sina Weibo data(Beijing area)at a specified time.These keywords are mainly divided into two categories:one is related to weather and climate,such as air,thunder,haze,etc.;the other is directly related to people’s willingness to travel,such as play,dinner party,work,etc.After acquiring these microblogs,in order to connect them with people’s travel intention,the paper selects two main features:The first is the number of microblogs with relevant keywords,and the second is the travel emotion polarity analysis of crawling microblogs with relevant keywords.Finally,according to the different crawling areas,the paper classifies them in space to reflect the travel intention of people in each area.Finally,these extracted features are output through the full connection layer and fused with the output of ST-LSTM-ResNet to get the final prediction results.The experimental results show that this method can get better prediction results than the existing five common methods on the same data set(8%).